v1.0.2

What's Changed

• Hotfix: update parameters that are reordered/concat in vllm parameter sync by @SeaOfOcean in #120

Full Changelog: v1.0.1...v1.0.2

v1.0.1

What's Changed

• [doc] refine performance image size by @SeaOfOcean in #39
• Update issue templates by @SeaOfOcean in #40
• Add unit test and daily build by @SeaOfOcean in #41
• raise error if parameter sync breaks by @charles9304 in #42
• Refine the description of docs. by @adoda in #45
• Trigger UT when the pull request is approved and the number is 2 by @SeaOfOcean in #52
• fix none src model and skip load ckpt for vllm by @stcy07 in #50
• refine log output for trainer model by @stcy07 in #49
• make *penalty of sampling_params configurable. by @charles9304 in #54
• set env concurrent. by @adoda in #55
• Refine vllm inference and keep the API same as non-vllm by @SeaOfOcean in #46
• Speedup ut && format. by @adoda in #57
• Refine efficient memory sharing by @SeaOfOcean in #58
• fix vllm_module InferenceMemoryManager args error by @SeaOfOcean in #61
• [UT] rm duplicate ray stop by @SeaOfOcean in #62
• fix onload offload in save_checkpoint by @SeaOfOcean in #63
• fix exit with log_monitor error by @SeaOfOcean in #60
• disable onload/offload when not colocated. by @charles9304 in #65
• Parameter sync fallback to P2P when TP size is odd by @SeaOfOcean in #64
• fix cpu_per_process and gpu_per_process when num_gpu/num_cpu is 1 by @SeaOfOcean in #67
• Reverse DP replicas in parameter sync when tp size is odd by @SeaOfOcean in #68
• Upload Python Package when release is published by @SeaOfOcean in #69
• stop previous run container when running ut by @SeaOfOcean in #73
• Support get tp/pp for torch_module/deepspeed_module and fix ut. by @adoda in #72
• Add DingTalk group to README. by @adoda in #74
• fix policy generation oom when continue train by @SeaOfOcean in #77
• Increase the num of episodes to allow the model to converge more fully by @adoda in #76
• set build time to 00.30 am utc+8 by @SeaOfOcean in #75
• feat:add and use multi thread tokenize tool in VLLMPromptPipeline by @stcy07 in #56
• add load ckpt for value model and warinings by @stcy07 in #78
• Be compatible with group query attention for QWen2. by @charles9304 in #79
• fix missing import in example by @SeaOfOcean in #80
• Upgrade version number by @SeaOfOcean in #81
• Revert "fix exit with log_monitor error (#60)" by @SeaOfOcean in #82
• fix dp_rank not in dp2send_actors when inference replica num less than training replica num by @SeaOfOcean in #83

New Contributors

• @charles9304 made their first contribution in #42
• @adoda made their first contribution in #45
• @stcy07 made their first contribution in #50

Full Changelog: v1.0.0...v1.0.1

v1.0.0

What's Changed

• Support vLLM as generation engine.
• Support custom flow.
• Support efficient memory sharing.
• Support CPU module.
• Add Llama2 DPO/OnlineDPO/GRPO example based on Megatron-LM.
• Add QWen2 DPO example based on DeepSpeed.

Full Changelog: v0.2.0...v1.0.0

v0.2.0

What's Changed

• Support llama2 based on official repo of Megatron-LM
• Refactor tutorial docs and add tools to convert megatron ckpt to hf
• Fix parameter sync when src_pipe != tgt_pipe and tgt_pipe != 1
• Reduce the number of port required
• Refine resume training and doc
• Add node address to error message and exit with error code
• Show the log in each worker node and refine docs
• Add continue train docs and check applied device
• Join log thread with timeout and trigger when process exit
• Support custom model flow
• Feat: support optimizer offload
• Doc: add faq

Full Changelog: v0.1.0...v0.2.0

v0.1.0

First release of ChatLearn

• Enable RLHF with Megatron-LM.
• Support Llama/GPT/Bloom/Baichuan models with SFT / Reward / RLHF.
• Add docs.
• Support resume training.